Inference cannot happen without training the model first, so the distinction is quite pointless.